N NVIDIA Last verified

NVIDIA H200 SXM 141GB

SXM 在售 发布于 2024 hopper-gen1
BF16
TFLOP/s
989 厂商声称
FP8
TFLOP/s
1979 厂商声称
FP4
TFLOP/s
不支持
Memory
GB
141 厂商声称
Mem BW
GB/s
4800 厂商声称
TDP
W
700 厂商声称

完整规格

算力

FP4 TFLOPS
不支持
FP8 TFLOPS
1979
BF16 TFLOPS
989
FP16 TFLOPS
989
INT8 TOPS
1979

显存

容量
141 GB
带宽
4800 GB/s
类型
HBM3e

芯片架构 🟢 vendor floorplan

SM count
132
Tensor cores / SM
4
L2 cache
50 MB
HBM stacks
6
制程
4 nm
Die area
814 mm²
Transistors
80 B
PCIe
Gen 5 ×16

Scale-Up (节点内)

协议
NVLink-4.0
单链带宽
900 GB/s
World size
8
拓扑
switched
交换机
nvswitch-gen3

Scale-Out (节点间)

单卡出口
400 Gbps
协议
InfiniBand-NDR
NIC
ConnectX-7

拓扑示意

拓扑结构 · Topology
8 卡 scale-up domain
芯片内部 / Die-level architecture
HBM HBM HBM HBM HBM HBM NVIDIA H200 SXM 141GB L2 / shared cache · NoC L1$ / register file (per SM) 132 SMs · darker block = tensor / matrix engine 989 TFLOPS BF16 · 1979 FP8 · 141 GB HBM3e @ 4.8 TB/s · 700 W TDP

🟢 vendor floorplan 132 SMs · 6× HBM · 50 MB L2 · 4 nm · 80 B transistors · 814 mm²


集群拓扑 / Cluster topology · NVLink-4.0 @ 900 GB/s
nvswitch-gen3 900 GB/s/link · all-to-all GPU 0 141GB GPU 1 141GB GPU 2 141GB GPU 3 141GB GPU 4 141GB GPU 5 141GB GPU 6 141GB GPU 7 141GB 8 cards · switched topology · scale-out: 400 Gbps/card
Scale-Up · 域内
NVLink-4.0
900 GB/s · 拓扑: switched
world_size = 8
Scale-Out · 跨域
InfiniBand-NDR
400 Gbps/卡 NIC
ConnectX-7

能跑哪些模型?

Quick estimates · decode tok/s/card 上界

TP=8 · FP8 · batch=16 · prefill=1024 · decode=256 · 已应用 efficiency 校准

在计算器中调整 →
模型 参数 (active) Decode tok/s/card 瓶颈
DeepSeek V4 Pro
deepseek
49B 显存不足
DeepSeek V4 Flash
deepseek
13B 409 内存带宽
Mistral Small 4
mistral
22B 187 内存带宽
GLM-5 Reasoning
zhipu
32B 154 内存带宽
GLM-5.1
zhipu
32B 105 内存带宽
Qwen3.6 Plus
alibaba
35B 100 内存带宽
Kimi K2.6
moonshot
32B 86 内存带宽
MiniMax M2.7
minimax
46B 68 内存带宽

算子级 fit · 任意模型瓶颈类型 + 上界

算子级 fit · operator-level fit (per-token roofline)

基于每个模型 operator_decomposition + 本卡 BF16 989 TFLOPS / 4,800 GB/s 计算 · ridge point ≈ 206 FLOPs/byte

上界 = min(计算屋顶, 内存带宽屋顶) · efficiency 未应用
模型 domain 主导算子 AI · F/B 瓶颈 tok/s 上界
DeepSeek V4 Pro llm matmul 245.5 🔥 计算 164k
GraphCast scientific graph-message-passing 0.9 💾 内存带宽 8856
AlphaFold 3 scientific pair-bias-attention 2.3 💾 内存带宽 2661
GPT-OSS llm matmul 0.7 💾 内存带宽 388
Gemma 4 26B llm matmul 0.7 💾 内存带宽 288
DeepSeek V4 Flash llm matmul 0.8 💾 内存带宽 273
Mistral Small 4 llm matmul 0.6 💾 内存带宽 124
Llama 4 Maverick llm matmul 0.8 💾 内存带宽 123
需要 efficiency 校准 + concurrency 扫描 + TCO 估算 → 在计算器中评估 →

算子支持 & 优化空间

算子支持 & 优化空间 / Operator support & headroom

Per-operator support derived from software_support.engines + scale-up topology. Optimization headroom from measured efficiency factor.

Optimization headroom
+-50 pp
saturated

Near saturation at 150% of roofline. Further gains require workload restructure (disaggregated, speculative, smaller batch).

Communication (collective)
All-to-All 🟢 mature
all-to-all via NVLink-4.0 world_size=8
AllReduce 🟢 mature
NVLink-4.0 ring all-reduce
Attention
Multi-Head Attention 🟢 mature
paged-attention via vLLM/SGLang/MindIE
FlashAttention-3 🟢 mature
FA-3 on modern engine + tensor cores
Matrix multiply (GEMM)
Matrix Multiplication 🟢 mature
GEMM supported on all inference engines
MoE routing
MoE Routing 🟢 mature
MoE gating supported via vLLM ≥0.4 / SGLang
Normalization
RMSNorm 🟢 mature
fused into engine kernels
Embedding
fused into engine kernels
Activation
SiLU / Swish 🟢 mature
fused into engine kernels
Softmax 🟢 mature
fused into engine kernels

软件栈支持

引擎 状态 BF16FP16FP4FP8 E4M3FP8 E5M2INT4 AWQ
HanGuangAI 未确认
LMDeploy 未确认
MindIE 未确认
MoRI 未确认
SGLang 官方
TensorRT-LLM (Dynamo) 官方
vLLM 官方
实测校准 efficiency factor

基于 2 个该硬件的实测案例计算得出, 计算器使用此值替代默认 0.5。

σ = 0.00 · range [1.50, 1.50]

1.50 ± 0.00
measured / theoretical (n=2)

已有部署案例 (2)

引证

  1. [1] NVIDIA H200 Tensor Core GPU product page — https://www.nvidia.com/en-us/data-center/h200/ · 访问于 2026-04-28 厂商声称
  2. [2] H200 reuses GH100 die (132 SMs, 50 MB L2, 814 mm²); 6× HBM3e stacks @ 24 GB ⇒ 141 GB capacity — https://resources.nvidia.com/en-us-tensor-core/gtc22-whitepaper-hopper · 访问于 2026-04-28 厂商声称
⚠ All performance figures are vendor-claimed unless tier=measured.